Surface defect detection of smartphone glass based on deep learning

Because of the high difficulty of smartphone glass detection and the variety of defect morphologies, the detection results are easily affected by the environment, making it difficult to meet the accuracy requirements of industrial inspection. Based on the existing YOLO v5s network, this study proposes a new network Dy-YOLO v5s. In particular, an attention module is introduced into the residual structure, and the cross-scale and cross-layer connections of feature maps are added to the Neck to improve the feature extraction and information exchange capabilities of the detection network. This algorithm introduces the dynamic detection framework called dynamic head (DyHead), which improves the detection head's capacity for perception. Additionally, the redundant anchor boxes and the balance of positive and negative samples are deduplicated using the confidence propagation cluster (cp-cluster) and varifocal loss functions. The experimental results demonstrate that when the intersection over union (IOU) threshold is set to 50%, the mean average precision (mAP) of Dy-YOLO v5s, precision rate (P), and recall rate (R) reach values of 96.2%, 92.6%, and 93.1%, respectively. Compared with YOLO v5s, mAP@0.5 and mAP@0.5-0.95 increased by 4.5% and 4.6%, respectively. The approach also has significant advantages over other deep-learning algorithms in terms of overall accuracy and real-time performance. Therefore, it can fully satisfy the detection requirements of smartphone glass.